autograph => AutoGraph

Done

Edit:
It's easy to use autograph with tf.keras, but batching is essential for performance. The best code to convert to AutoGraph is code where the control flow is decided at the batch level. If making decisions at the individual example level, you must index and batch the examples to maintain performance while applying the control flow logic.

I suggest keeping the observation about batching as a separate callout, e.g. in a separate paragraph with "Tip: remember that the inputs are usually batched examples, ...". The guidance is otherwise not strongly related to the keras interop.

Edit:

Done.

not strongly related to the keras interop.

Moved up to the of examples.

For stateless functions, like collatz shown below, the easiest way to include them in a Keras model is to wrap them up as a layer using tf.keras.layers.Lambda:

Done

TODO now or later?

When I can show that this performs as the paper: https://colab.research.google.com/drive/14YkY78YjBhqyKn9pQ50rC36hWkuWHFb7#scrollTo=nzst0_CH2uXV

The easiest way to use AutoGraph with Keras layers and models is to @autograph.convert() the call method. See the TensorFlow Keras guide for details to build these classes.

Done

Use a DevSite callout, options are:
Note:, Caution:, Warning:, Key Point:, etc

!

I was sure "important" was one of the allowed ones, maybe that was in g3doc....

Fixed/reworded.

Not related to this PR, but I feel that we should avoid examples that place control flow outside sess.run - in a way, AutoGraph was made so that one doesn't have to use that, which is a bit of an antipattern. For the point above, that converted code works like a regular op, I think we can use anything, like for instance taking the output and passing it through some other op, so we don't strictly need the fizzbuzz example for that. LMK what you think.

I see your point.

Here I'm trying to show that FizzBuzz is working, without getting into too-complex of an example.

This would be a step in the right direction:

@autograph.convert()
def fizzbuzz(num):
  if num % 3 == 0 and num % 5 == 0:
      return 'FizzBuzz'
  elif num % 3 == 0:
      return 'Fizz'
  elif num % 5 == 0:
      return 'Buzz'
  else:
      return tf.as_string(num)
    
@autograph.convert()
def fizzbuzz_loop(start, stop):
  for n in range(start, stop):
    print(fizzbuzz(n))
  return tf.constant(stop)

with tf.Graph().as_default():
  # The result works like a regular op: takes tensors in, returns tensors.
  # You can inspect the graph using tf.get_default_graph().as_graph_def()
  with tf.Session() as sess:
    sess.run(fizzbuzz_loop(10,16))

But print is currently not working...

Style nit: for expressions, the Google style guide recommends placing a space around operators, e.g. "1.0 / x"; I think it would be useful to be consistent with it in this notebook.

+1 we just don't have an integrated notebook linter (yet?).

Done.

could you use sess.run(inverse(...)) instead of .eval() here and elsewhere?

done

I suggest keeping the observation about batching as a separate callout, e.g. in a separate paragraph with "Tip: remember that the inputs are usually batched examples, ...". The guidance is otherwise not strongly related to the keras interop.

FYI, we're looking into modifying the default behavior to avoid the need for these explicit casts.

+1, good to know.

Another way this could be framed is that AutoGraph is compatible with Keras: AutoGraph code can be used inside Keras models and layers as well; Keras models can also be used in AutoGrahp code, as this example shows.

reworded, wdyt?